realrep

DeCoMix-HDR is the official implementation of a deep-learning method for generalized SDR-to-HDR conversion. It is designed to handle SDR inputs produced by varied tone-mapping and degradation styles, including different tone curves, exposure shifts, highlight compression, saturation changes, and color casts. The method learns a shared attribute-disentangled degradation representation using luminance- and chrominance-aware contrastive supervision. A UNet-based encoder produces global and local degradation descriptors. Candidate negatives are collected from SDR versions of the same HDR scene generated with different degradation styles, then ranked using log-luminance distance and Yxy chromaticity distance. Rank-guided representation mixing interpolates selected negatives toward positive keys to create continuous hard negatives. The Degradation-Adaptive HDR Mapping network uses dense global channel modulation and sparse local pixel-wise modulation to reconstruct HDR output. Negative samples and the momentum encoder are used only during training; inference requires one SDR image and a trained checkpoint. The repository follows a BasicSR-style structure and includes training and evaluation entry points, YAML configuration files, model architectures, contrastive dataset loaders, loss and metric utilities, distributed launch scripts, and dataset preparation helpers. A paper-to-code mapping identifies the training wrapper, DAHM architecture, contrastive encoder, ranked-negative dataset, latent mixing routine, staged training logic, and checkpoint keys. The documented software environment uses Python 3.8 or newer, PyTorch with CUDA, torchvision, kornia, lmdb, and OpenCV. The main experiments use HDRTV4K or HDRTVDM for training and known-degradation evaluation, and HDRTV1K for unknown-degradation evaluation. Datasets are expected in LMDB format with name metadata and ranked luma and chroma negative-list files. A helper script can create dataset symlinks and local configuration files and verify required data. The README provides single-GPU and distributed training commands, a default 400,000-iteration schedule, checkpoint paths, and full-validation commands. Evaluation outputs include PSNR, SSIM, and Delta E ITP measurements. The results table also reports HDR-VDP3. Representative reported results are 31.20 PSNR and 0.9213 SSIM on HDRTV4K known degradations, and 29.12 PSNR and 0.9412 SSIM on HDRTV1K unknown degradation, with the README noting that exact values depend on dataset preparation, checkpoint, and evaluation environment. Some historical RealRep class and file names remain for compatibility with model registries and checkpoints. Ablation configurations, training logs, paper sources, figures, PDFs, local datasets, and experiment outputs are not included in the public package.